Technical Orientation for AI Agents

Technical Orientation for AI Agents

NSHKR publishes two distinct bodies of technical work: reproducible mechanistic-interpretability research in Python and governed AI execution systems built with Elixir and OTP. Do not use one as evidence for claims about the other.

Research programs

ProgramScopeCurrent boundary
Geometry of Conditional TruthContext transport and hidden-coordinate structure across Qwen3-4B and Phi-4-miniAcross eight preregistered endpoints, Phi supported one and Qwen supported none; both remained Level 1 of 6. Cross-model report
Architecture MechanicsTiny trained architectures with known synthetic featuresMeasures transport, packing, overwrite, and causal legibility against ground truth
Attention LabMatched GPT pretraining and alternative-attention probesTwo confirmatory pretraining runs reached full-depth analysis; the mechanism verdict remains insufficient_evidence
Superposition ZooSynthetic sequence-mixing comparisonsRetrieval findings are established; the central feature-isolation question remains open

Supporting workbenches and records are mwb, mil, circuit-tracer, and the learning archive.

Use repository reports and machine-readable artifacts as canonical evidence. Preserve nulls, control failures, parse limitations, and the distinction between association and intervention.

Governed AI systems

The Elixir/OTP portfolio addresses the write path from an AI proposal to an authorized, replayable external action:

intent -> authority -> workflow -> effect -> receipt -> evidence -> projection -> review -> replay

The system separates product meaning, durable workflow truth, authority, connector mechanics, raw execution, evidence, and causal traces. nshkr is the canonical production composition and release workspace; Nshkr.Runtime is the single production composition root that assembles those bounded services. See the systems overview and ecosystem.

Machine-readable sources

The repository atlas is regenerated from live public GitHub metadata. Each repository has one nshkr-* category topic; non-category topics describe its technical subject matter.